optimize branch logic

[mcroomp]

based on feedback from #49 from @Melirius

Analysing the code I got to conclusion that corner case even if present in the initial code of Lepton, does not affect its behaviour. Citing https://github.com/microsoft/lepton_jpeg_rust/blob/main/src/structs/branch.rs

The only corner case is that in the case of 255 true and 1
false, the C++ version decides to set the probability to 0 for the next
true value, which is different than the formula ((a << 8) / ( a + b )).
The formula produces 1 here, and this does not affect neither https://github.com/microsoft/lepton_jpeg_rust/blob/main/src/structs/vpx_bool_reader.rs#L147 nor https://github.com/microsoft/lepton_jpeg_rust/blob/main/src/structs/vpx_bool_writer.rs#L159 as from VP8 spec range of tmp_range is [128,255], so that both bugged and formula version give split = 1.

[mcroomp]

@danielrh @Melirius found a way to make this even faster

[danielrh]

This is so cool! After all these years you found a nice win in the inner loop :-)